planning-with-crux

name

planning-with-crux

description

Design and author agentic workflows using the crux DSL — both YAML pipeline files (.crux) and Rust macro-based agents. Use when writing a .crux pipeline, designing an agent workflow, choosing between crux combinators (pipe, join_all, speculate, route_on_confidence, delegate), or planning how to decompose a task into crux steps. Triggers on "write a pipeline", "crux pipeline", ".crux file", "agentic workflow", "agent plan", "step handler", "which combinator", or any request involving pipeline design, step composition, or handler selection. Also use when reviewing or debugging existing pipelines.

Planning with Crux

Crux makes agentic control flow explicit. Every step, branch, delegation, and failure is a first-class value in the trace. This skill helps you design workflows that use these primitives effectively.

Two Modes of Authoring

1. YAML pipelines (.crux files) — declarative, no Rust compilation needed. Run with crux run pipeline.crux [input.json]. Good for orchestration tasks that compose built-in handlers.

2. Rust agents (#[crux::agent]) — full type safety, custom logic, async. Compile into binaries. Good for complex domain logic that needs Rust's type system and error handling.

Choose YAML when the workflow is a composition of existing handlers (shell commands, file ops, git, JSON transforms, LLM calls). Choose Rust when you need custom types, complex branching, or the workflow will be a library.

Pipeline Anatomy

pipeline: my_workflow # Required: name (appears in trace)
budget: { calls: 10 } # Optional: tokens, calls, duration_ms, cost_cents
steps: # Required: list of step definitions
  - step: first_step
    handler: shell::capture
    args:
      cmd: "echo hello"

Every pipeline produces a Crux<Value> trace. The last step's output is the pipeline result.

Choosing a Combinator

Pattern	Combinator	When to use
Do A, then B, then C	pipe:	Sequential stages, each feeds next
Do A, B, C simultaneously	join_all:	Independent work, merge results
Try multiple approaches	speculate: first_ok	Fallback chain, first success wins
Pick the best attempt	speculate: pick_best	Race approaches, score and select
Branch on confidence	route_on_confidence:	Different actions for different scores
Hand off to sub-agent	delegate:	Encapsulated child agent
Single action	step:	One handler invocation

Read references/combinator-patterns.md for worked examples of each pattern and guidance on composing them together.

Handler Selection

Read references/handler-catalog.md for the complete handler reference organized by category (shell, fs, git, json, text, llm, ci, review, etc.).

Key principles:

• shell::capture fails on non-zero exit; shell::exec ignores it. Use capture when the command must succeed. Use exec for best-effort.
• handler() vs handler_value(): If a handler needs to emit confidence for route_on_confidence, register with handler() and return HandlerOutput::with_confidence(value, score). If it just returns data, use handler_value().
• Template expressions {{ steps.X.output.field }} and {{ input.field }} expand in args: values. Use these to pass data between steps.

Budget Design

Set budgets to prevent runaway execution:

budget: { calls: 10 }             # Max 10 handler invocations
budget: { tokens: 15000 }         # Max 15k tokens (LLM steps)
budget: { duration_ms: 300000 }   # Max 5 minutes wall clock
budget: { cost_cents: 50 }        # Max $0.50

For Rust agents, budgets can be scoped per delegation:

x.delegate::<SubAgent>("task", input)
    .with_budget(Budget::tokens(5000))
    .run()
    .await?

Pipeline Design Process

1. Name the outcome — what does the pipeline produce?
2. List the steps — what information or actions are needed?
3. Identify dependencies — which steps need outputs from other steps?
4. Choose combinators — sequential (pipe), parallel (join_all), or branching?
5. Set the budget — what's the maximum acceptable cost/time?
6. Add confidence routing — should different confidence levels trigger different actions?

Rust Agent Patterns

See references/rust-agent-patterns.md for the #[crux::agent] macro patterns, including steps, delegation, speculation, lifecycle hooks, replay, and task registry integration.

Common Mistakes

• Forgetting handler: on a step — the step name is used as handler name if handler: is omitted. This works if your handler is named the same as your step, but fails silently if not.
• Non-covering confidence routes — ranges must cover [0.0, 1.0] with no gaps. Use [0.0, 0.5), [0.5, 0.8), [0.8, 1.0] — note the inclusive end on the last range.
• Using handler_value() with route_on_confidence — handler_value() handlers don't emit confidence. Use handler() and return HandlerOutput::with_confidence().
• speculate: pick_best without score — arms must include a score field in their output JSON. Arms without score all tie at 0.0.